/06
Writing
Deep dives on things I built and what broke along the way, plus close-read notes on the papers behind them.
Long-form
- Jun 17, 2026
VLA-DSS: a Vision-Language-Action model on the efficiency frontier
How a 28.9M-parameter, RGB-only robot policy holds its own against models several times its size, by trading learned visual capacity for mathematically-grounded structure: wavelet scattering, a frozen DINOv3 backbone, and a Fourier Neural Operator action head.
- Apr 03, 2026
The Axiom Journey: Engineering a High-Performance Engine
A technical dive into the core architecture, sprite batching, and Entity-Component Systems (ECS) of the Axiom 2D game engine.
Paper notes
Close reads and worked-through math, the things I wanted to remember exactly, kept by field as the work shifts.
Vision-language-action & robot learning 9
- May 08, 2026
A survey of Vision-Language-Action models
Orientation notes on the VLA landscape — pretrained visual representations (CLIP, MAE), control policies (tokenization vs diffusion), and the dataset problems that open datasets set out to fix.
- May 06, 2026
What Matters in Building a VLA
The core design axes when building a VLA — discrete vs continuous actions, where history enters, diffusion vs MSE action generation, OXE pretraining, and the mixture-of-experts trade-off from π₀.
- May 02, 2026
BitVLA — 1-bit Vision-Language-Action models
Pushing VLA weights down to {0, ±1}: why naive ternary weights collapse the model, the distillation trick that keeps a quantized vision encoder usable, and the latency/throughput it buys.
- Apr 25, 2026
HAMLET — turning a VLA into a history-aware policy
Robot manipulation is non-Markovian, but feeding raw past frames causes causal confusion. HAMLET compresses each timestep into moment tokens and reads them with a transformer memory module — near-free latency, big real-world gains.
- Apr 18, 2026
COT Policy — fast flow via conditional optimal transport
Straightening flow-matching paths during training itself — no distillation — by choosing which noise pairs with which action via condition-aware optimal transport. Two-step COT beats 20-step Diffusion Policy.
- Apr 11, 2026
KAN We Flow — an RWKV + KAN flow-matching head
Replacing the heavy UNet in a flow-matching policy with RWKV time-mixing and GroupKAN channel-mixing: 86.8% fewer parameters, ~10x faster inference, and still the best accuracy across difficulty tiers.
- Apr 04, 2026
π₀ — a flow-matching VLA for dexterous control
How π₀ pairs a frozen PaliGemma VLM with a from-scratch flow-matching action expert for 50 Hz dexterous control, why vanilla flow matching is multi-step, and the KV-cache trick that makes it fast.
- Mar 28, 2026
Octo's diffusion action head, line by line
A from-first-principles deep dive into Octo's DiffusionActionHead — DDPM math, the cosine schedule, the score network, and every line of the training and inference code — and exactly what a flow-matching head would replace.
- Mar 21, 2026
Octo — an open-source generalist robot policy
Octo's token-based modular design: how a generalist policy pretrains on 800k trajectories and finetunes to a new robot in hours by swapping tokenizers and a diffusion action head while the transformer stays frozen.